Machine learning overview

2
Overview

� we describe here the general ideas and methods used in machine learning, at a high level

� we will go over all of these topics later, in much more detail

� don’t worry if some of this seems abstract at this point, or there are terms you don’t know yet

3
Artificial intelligence approaches

� we would like computers to perform complicated tasks, e.g., medical diagnosis

� distinguish two approaches

� knowledge-based: a computer program whose logic encodes a large number of properties of the world,
usually developed by a team of experts over many years
� machine learning: extract information directly from historical data and extrapolate to make predictions

� this class is about machine learning

4
Machine learning tasks

generic tasks:

� build a model from some data

� choose how to map raw data to feature vectors
� choose a model form
� choose parameter values in the model

� test or validate the model

� evaluate the model on unseen data to assess its performance

‘model’ can mean several things, depending on context

5
Machine learning model taxonomy: supervised vs. unsupervised

� supervised learning models predict something, given some other things

� called a prediction model
� called regression when we predict a real scalar or vector value
� called classification when we predict a value from a finite set such as �True� False�
� called forecasting when you predict a future value, given current and past values

� unsupervised learning models just create a model of the data

� called a data model

� the lines between these can be blurred

6
Machine learning model taxonomy: point vs. probability

� a point estimate predicts a single value

� a probability estimate predicts a distribution of values

� a confidence band estimate predicts a confidence band or interval of values

� a generative model generates samples from an estimated distribution

� the lines between these can be blurred

7
Examples

what kind of models would each of these tasks use?

� predict tomorrow’s rainfall, given the date and the last 10 days of rainfall data

� determine from a photo of a face if the user is who she claims to be

� estimate the probability of 10 possible diagnoses, given some patient data, test results

� cluster customers into 22 different groups with similar buying habits

� estimate the risk (probability) of an auto accident at a location given the hour and weather

� build an anomaly detector, that rates how suspicious some new data is

� build a simulator that generates fake new data that ‘looks like’ the given data

� build a recommendation engine that suggests products a customer might be interested in

8
Performance metrics

� we judge performance of a model on some data using a metric such as

� mean-square or RMS prediction error (for regression)
� error rate (for classification)
� log likelihood (for probabilistic models)

� examples:
� our predictor predicts tomorrow’s maximum temperature with an RMS error of ���� C
� our classifier predicts the topic of a newspaper article (from a set of 50 choices) with an error rate of ��

9
Training and validation

� our goal is to develop a model that performs well for new, unseen data

� standard practice is to divide the given data set into two parts
� a training data set, used to choose or train the model
� a test or validation data set, used to evaluate tentative models

� we can look at the performance metrics on the training and test data sets

� if the model performs well on the training set, but poorly on the test set, it is overfit
(and probably useless in practice)

� if the model performs well on the test set, it’s likely going to perform well on new unseen data

10
Learning a model

a common method of choosing a model:

� choose the model structure or form or type

� the model contains a number of model parameters

� choose a loss function that rates how badly the model performs on a single data point or example

� choose the parameter value by minimizing an average loss over the training data

this general scheme, called empirical risk minimization, is used to fit a wide variety of models

11
Examples

12
Example: diagnosis

� goal is to predict if a patient has a disease, based on whether or not she exhibits 10 symptoms

� historical data consists of a large number of patient records

� each record contains

� 10 Booleans, specifying the presence or absence of the 10 symptoms
� a Boolean specifying whether that patient had the disease

� machine learning algorithm observes these data, produces a predictor

� predictor takes as input 10 Booleans, returns a single Boolean prediction

� this is a classifier, since we are predicting an outcome that takes only two values

� we will judge model by its error rate on a separate test set of data, not used to develop the model

� a probabilistic model returns a probability that the patient has the disease, not just a Boolean

13
Example: digit recognition

� 8-bit grayscale images of handwritten digits, �� � �� pixels

� goal is to guess the digit (i.e., �� � � � � �) from the image

� 60,000 training images, 10,000 test images

� preprocessed by antialiasing, scaling and centering

� data originally by NIST, modified by Le Cun, 1998

14
Data

� Kaggle: datasets and competitions

� ImageNet dataset: 14m images

� Street view house numbers: 600,000 digit images

� Waymo open dataset: self-driving car data

� many other large datasets

15
Software

� PyTorch, Keras, TensorFlow, JAX, and many others, . . .

� Scikit-Learn, Spark/MLlib

� Flux, in Julia

� generic packages in R, Matlab, Python, Java, Julia . . .

16
Course outline

17
The major parts of the course

1. regression — to predict one or more real values

2. classification — to predict one of a finite number of possible outcomes

3. probabilistic supervised learning — to predict a distribution of outcomes

4. unsupervised learning — to develop a data model

5. optimization — to fit or choose parameters in all of the models above

18